ZeroDDI: A Zero-Shot Drug-Drug Interaction Event Prediction Method with Semantic Enhanced Learning and Dual-Modal Uniform Alignment

TL;DR

This work addresses zero-shot DDIE prediction by learning biologically grounded DDIE representations and aligning drug-pair and DDIE semantics on a unit sphere. It introduces BRL to fuse class-level and attribute-level semantics with substructure-guided discrimination and DUA to mitigate class imbalance through contrastive alignment and dual-uniformity losses. Empirical results on DrugBank-derived CZSL and GZSL benchmarks show substantial improvements over baselines, and a novel dataset analysis confirms transferability to unseen DDIEs. The approach offers a practical, interpretable framework for detecting unseen DDIEs and provides insights into semantic-token–substructure relationships.

Abstract

Drug-drug interactions (DDIs) can result in various pharmacological changes, which can be categorized into different classes known as DDI events (DDIEs). In recent years, previously unobserved/unseen DDIEs have been emerging, posing a new classification task when unseen classes have no labelled instances in the training stage, which is formulated as a zero-shot DDIE prediction (ZS-DDIE) task. However, existing computational methods are not directly applicable to ZS-DDIE, which has two primary challenges: obtaining suitable DDIE representations and handling the class imbalance issue. To overcome these challenges, we propose a novel method named ZeroDDI for the ZS-DDIE task. Specifically, we design a biological semantic enhanced DDIE representation learning module, which emphasizes the key biological semantics and distills discriminative molecular substructure-related semantics for DDIE representation learning. Furthermore, we propose a dual-modal uniform alignment strategy to distribute drug pair representations and DDIE semantic representations uniformly in a unit sphere and align the matched ones, which can mitigate the issue of class imbalance. Extensive experiments showed that ZeroDDI surpasses the baselines and indicate that it is a promising tool for detecting unseen DDIEs. Our code has been released in https://github.com/wzy-Sarah/ZeroDDI.

Figures (8)

• Figure 1: (a) DrugBank is a standard database containing the textual descriptions of DDIEs. Almost every description can be split into Effect, Sign, and Pattern attributes. (b) With the updates of the Drugbank version, the number of DDIEs is also increasing. (c) Unseen DDIEs may come from the composition of existing attributes.
• Figure 2: The overall framework of ZeroDDI.
• Figure 3: Performance (in %) comparisons of ZeroDDI with its variants in the CZSL and GZSL scenarios.
• Figure 4: The visualization of an example of DDIE textual description with its corresponding drug pair molecular structures. The attention scores between the word "hypertensive" and substructures are highlighted in red colour.
• Figure 5: The visualization of drug pair representation distribution of test unseen and seen classes in GZSL scenario. Class Center denotes the center of all instances in a class. $\rho$ denotes imbalance ratios of training data. The $\textbf{acc}_{ave}$ here is the average accuracy of five DDIE classes.